Traffic Sign Classification (TSC) is crucial for autonomous driving and intelligent transportation systems. Desktop implementations of deep learning achieved state-of-the-art performance on TSC benchmarks; however, they are unsuitable for real-time embedded systems due to resource limitations. We propose an Efficient GPU-Embedded Network (EGENet) for embedded platforms, such as NVIDIA’s Jetson, to overcome these drawbacks. When implemented on a desktop system with NVIDIA GeForce RTX 2080, EGENet can reduce the number of parameters by 24 million while speeding up by 2.59 \(\times \) . EGENet introduces a new concept called Asymmetric Depth Dilated Separable Convolution (ADDSC), which enables a reduction in parameters and inference time while maintaining the receptive window size. A novel evaluation metric is proposed, considering frames per second (FPS), accuracy, and deployment on embedded GPU devices with constrained resources, targeting at least 98.85% accuracy and a frame rate of more than 30 FPS. Thorough evaluations were performed on the NVIDIA Jetson Xavier AGX and Jetson Nano, utilizing limited resources, to validate EGENet’s real-time performance. Evaluation of GTSRB and LISAC datasets demonstrates outperforming results, with an accuracy of 99.58% and 98.18% and a response time of 253 FPS and 90 FPS on Jetson Xavier AGX and Jetson Nano devices, respectively. Our work contributes to efficient TSC systems based on embedded GPUs and offers a comprehensive performance evaluation methodology for autonomous driving. We present exhaustive statistical comparative tests against state-of-the-art systems.
